NBER WORKING PAPER SERIES LABOR MARKET FLOWS IN THE CROSS SECTION AND OVER TIME. Steven J. Davis Jason Faberman John C.

Transcription

1 NBER WORKING PAPER SERIES LABOR MARKET FLOWS IN THE CROSS SECTION AND OVER TIME Steven J. Davis Jason Faberman John C. Haltiwanger Working Paper NATIONAL BUREAU OF ECONOMIC RESEARCH 1050 Massachusetts Avenue Cambridge, MA August 2011 This paper was presented at the Spring 2011 Carnegie-Rochester Conference. We thank Jennifer Hayden for excellent research assistance. We thank seminar participants at EUI, Princeton, Chicago, Rochester, the Federal Reserve Bank of Philadelphia, Federal Reserve Bank of New York, workshop participants at various conferences and Ay egül ahin, Hermann Gartner, Christopher Reicher, Tom Cooley and Yongsung Chang for comments on an earlier draft. We thank the Kauffman Foundation and the Booth School of Business at the University of Chicago for financial support. The views expressed are solely those of the authors and do not necessarily reflect the official positions or policies of the Federal Reserve Bank of Philadelphia, the Federal Reserve System, the U.S. Bureau of Labor Statistics, the U.S. Bureau of the Census, the National Bureau of Economic Research, or the views of other staff members. NBER working papers are circulated for discussion and comment purposes. They have not been peerreviewed or been subject to the review by the NBER Board of Directors that accompanies official NBER publications by Steven J. Davis, Jason Faberman, and John C. Haltiwanger. All rights reserved. Short sections of text, not to exceed two paragraphs, may be quoted without explicit permission provided that full credit, including notice, is given to the source.

2 Labor Market Flows in the Cross Section and Over Time Steven J. Davis, Jason Faberman, and John C. Haltiwanger NBER Working Paper No August 2011 JEL No. E24,J63,J64 ABSTRACT Many theoretical models of labor market search imply a tight link between worker flows (hires and separations) and job gains and losses at the employer level. Partly motivated by these theories, we exploit establishment-level data from U.S. sources to study the relationship between worker flows and job flows in the cross section and over time. We document strong, highly nonlinear relationships of hiring, quit and layoff rates to employer growth in the cross section. Simple statistical models that capture these cross-sectional relationships greatly improve our ability to account for fluctuations in aggregate worker flows. We also evaluate how well various theoretical models and views fit the patterns in the data. Aggregate fluctuations in layoffs are well captured by micro specifications that impose a tight cross-sectional link between worker flows and job flows. Aggregate fluctuations in quits are not. Instead, quit rates rise and fall with booms and recessions across the distribution of establishment growth rates, but more so at shrinking employers. Finally, we use our preferred statistical models in combination with data on the cross-sectional distribution of establishment growth rates to construct synthetic JOLTS-type measures of hires, separations, quits and layoffs back to Steven J. Davis Booth School of Business The University of Chicago 5807 South Woodlawn Avenue Chicago, IL and NBER Steven.Davis@ChicagoBooth.edu John C. Haltiwanger Department of Economics University of Maryland College Park, MD and NBER haltiwan@econ.umd.edu Jason Faberman Economic Research Department Federal Reserve Bank of Chicago 230 S. LaSalle St. Chicago, IL jfaberman@frbchi.org An online appendix is available at:

3 1. Introduction Many theoretical models of labor market search carry strong implications for the relationship of hires and separations to job gains and losses at the employer level. Partly motivated by these theories, we exploit establishment-level data from U.S. sources to study the empirical relationship between worker flows and job flows in the cross section and over time. The evidence supports a hybrid view that incorporates aspects of several distinct theories of labor market dynamics. Tracking the cross-sectional distribution of establishment growth rates, and applying simple models of cross-sectional behavior, greatly improves our ability to account for aggregate fluctuations in hires, separations, quits and layoffs. Previous work has not delivered a thorough and convincing explanation for the relationship between worker flows and job flows. One difficulty is that worker flow and job flow measures typically derive from different data sources, and are comparable only at high levels of aggregation. Even then, standard sources of data on these measures differ in scope, sampling frequency and other respects that hinder direct comparison. We overcome these difficulties by exploiting micro data from the Job Openings and Labor Turnover Survey (JOLTS). JOLTS data yield internally consistent measures of hires, separations, quits, layoffs, job creation and job destruction at the establishment and aggregate levels. To deal with weaknesses in the JOLTS sample design, we rely on the comprehensive Business Employment Dynamics (BED) to track the cross-sectional distribution of establishment-level growth rates over time. Specifically, we combine BED data on the establishment growth rate distribution with JOLTS data for hires, separations, layoffs and quits conditional on establishment growth rates to measure the aggregate worker flow rates. Combining JOLTS and BED data also enables us to construct synthetic data on worker flows back to 1990, nearly doubling the time-series length of JOLTS-type worker flow measures. 1

4 To guide our empirical work, we consider several theoretical models. Models in the spirit of seminal work by Mortensen and Pissarides (1994) (hereafter MP) imply a tight link between job flows and worker flows. Our empirical work assesses how closely models with these tight links fit the data. We also consider empirical specifications motivated by models that stress the role of search on the job, learning about match quality, firm profitability, and the impact of aggregate conditions on the propensity to quit. We evaluate alternative specifications in terms of their fit to cross-sectional patterns in the establishment-level data, their ability to generate the observed movements in aggregate worker flows, and their marginal explanatory power for aggregate worker flows after conditioning on aggregate variables. Figures 1 and 2 plot our quarterly, seasonally adjusted time series for JOLTS-type worker flow measures, along with job creation and destruction rates from the BED. 1 All rates cover the nonfarm private sector and are expressed as a percent of employment. We define job creation as the sum of employment gains at new and expanding establishments as in Davis, Haltiwanger, and Schuh (1996). We define job destruction analogously. Figure 1 shows that job destruction and layoffs move together over time, while quits move counter to both. Job destruction and layoffs exhibit pronounced spikes in late 2008 and early 2009 and then decline in nearly parallel fashion. Figure 2 shows that job creation and hires rates decline from 2006, bottom out in 2009Q1, and then partly recover. The hiring rate moves much more than the job creation rate over this period. Our study explores the micro level sources of these aggregate movements. Looking across establishments, hiring and separation rates exhibit powerful elements of the iron-link behavior implied by MP-style search models. That is, hires are tightly linked to job creation, 1 To conform to the BED sampling frequency, we cumulate monthly establishment-level JOLTS observations to the quarterly frequency. We then construct our aggregate worker flow series by combining the cross-sectional relationships of worker flows to establishment-level growth rates in JOLTS micro data with employment growth rate distributions from the BED. This approach, described more fully in Section III below, follows Davis, Faberman, Haltiwanger and Rucker (2010). Worker flows prior to 2001Q3 in Figures 1 and 2 are synthetic data, constructed as described in Section V below. 2

5 and separations are tightly linked to job destruction. When plotted as functions of establishment-level growth rates, hiring and separation rates exhibit nonlinear hockey-stick shapes. The hires relation is nearly flat to the left of zero growth (contracting employers) and rises more than one-for-one with employment growth to the right of zero (expanding employers), with a pronounced kink at zero. The separations relation is a mirror image of the hires relation. The hockey-stick shape for separations is starkly at odds with the simplifying assumption adopted in many search models of a uniform separation rate. Turning to the components of separations, both quits and layoffs rise with job destruction, but layoffs dominate the adjustment margin for rapidly contracting establishments. We also consider how the cross-sectional relations vary with aggregate conditions. As it turns out, the cross-sectional layoff relation is highly stable over time. In contrast, the cross-sectional quit relation varies markedly with aggregate conditions. Specifically, the quit relation shifts downward when aggregate conditions are weak, especially at contracting establishments. Several theoretical models provide insight into the reasons for departures from an iron-link relationship between worker flows and job flows. Faberman and Nagypál (2008) consider a model of on-the-job search that delivers an abandon-ship effect at struggling employers. Firms vary in their idiosyncratic profitability and grow faster when more profitable. Because wages increase with firm profits in their model, workers at low-profitability firms are more likely to accept an outside offer. Consequently, quit rates decline with the idiosyncratic component of firm profitability. Barlevy (2002) considers a model with on-the-job search and a match-specific component of productivity. Because firms create fewer vacancies when aggregate conditions are weak, employed workers encounter better matches at a slower pace during recessions. As a result, workers tend to remain longer in poor matches, which Barlevy refers to as the sullying effect of recessions. In sum, Faberman and Nagypál highlight the role of cross-sectional variation in firm-level circumstances for quit rates, whereas Barlevy 3

6 highlights the role of variation over time in aggregate conditions. Our evidence indicates that both effects are at work. In Jovanovic (1979, 1985) and Moscarini (2005), gradual learning about match quality leads to a separation rate that declines with match tenure. Because more rapid growth involves a higher share of young matches, these models suggest that separations rise with the growth rate of expanding employers. Pries and Rogerson (2005) integrate elements of Jovanovic-style learning into an MP model. Separations occur because of job destruction, as in the MP model, and because of learning effects about match quality. Thus, the model of Pries and Rogerson generates elements of iron-link behavior in hires and separations while rationalizing a positive relationship between separations and growth at expanding employers. The data support this hybrid view of the cross-sectional relationship between hires, separations and employer growth. Motivated by these theoretical ideas, we develop parsimonious statistical models of how worker flows vary in the cross section, and how the cross-sectional relations move over time. The statistical models serve three objectives. First, they provide guidance in evaluating, developing and calibrating theoretical models of labor market flows. Second, they allow us to investigate whether tracking the cross-sectional growth distribution adds to our understanding of aggregate movements in labor market flows. Third, they yield a framework for constructing synthetic data on aggregate hires, separations, quits and layoffs in the period before the advent of JOLTS. When we consider specifications that impose time-invariant relations of worker flows to employer growth rates in the cross section, the statistical models perform reasonably well in tracking the aggregate movements of hiring and layoff rates. However, the same type of model fails miserably in tracking the aggregate behavior of quits. Consequently, it also fares poorly in accounting for the 4

7 aggregate separations rate. When we allow the worker flow-employer growth relationships to vary with aggregate conditions, our ability to track aggregate quits and separations improves dramatically. Our work in this paper has many antecedents. There is a large body of previous research on job flows and worker flows. We review research in this area in Davis and Haltiwanger (1999) and Davis, Faberman and Haltiwanger (2006). Labor market flows and job vacancies play central roles in modern theories of unemployment based on search and matching models. See, for example, Pissarides (2000), Rogerson, Shimer and Wright (2005) and Yashiv (2007) for reviews of work in this area. Models that treat hires as the outcome of a matching function carry implications for the relationship between hires and vacancies in the cross section and over time. We explore some of those implications in Davis, Faberman and Haltiwanger (2010). The paper proceeds as follows. Section II discusses the conceptual underpinnings that guide our empirical work. We start with the model of Cooper, Haltiwanger and Willis (2006), which extends the basic MP model to multi-worker firms. We then consider models that endogenize the worker s quit decision, as in Faberman and Nagypál (2008) and Barlevy (2002), and conclude with models of learning about match quality, such as Jovanovic (1979). Section III describes our data and empirical measures. Section IV presents our statistical models and investigates how well they account for worker flows in the cross section and over time. Section V constructs synthetic JOLTS-type data on worker flows, and Section VI concludes. 2. Conceptual Underpinnings and Theoretical Implications In thinking about worker flows and job flows, it is useful to begin with an identity: e it = eit 1 + h it l it q it, (1) where e it is employment at establishment i in period t and h, l and q denote hires, layoffs and quits. Separations are the sum of quits and layoffs. Theory provides guidance about how employers use hires, 5

8 layoffs and quits to adjust employment and about the factors that lead to worker turnover in excess of employment changes. In what follows in this section, we first consider search and matching models that involve an iron link between hires and job creation on the one hand and separations and job destruction on the other. We then consider theories that relax the iron link. Lastly, we discuss aggregate implications and motivate our empirical specifications Models with an Iron Link Every hire reflects a newly created job in the canonical search and matching model of Mortensen and Pissarides (1994), and every separation reflects a job that vanishes. That is what we mean by an iron link between worker flows and job flows. One goal of our study is to assess how well this iron link characterizes the data. However, the basic MP model has no role for multi-worker employers, an essential aspect of our empirical work. So we borrow from Cooper, Haltiwanger and Willis (2007, CHW) to illustrate the implications of the iron link in a multi-worker version of MP. 2 Employers in the CHW model face common and idiosyncratic shocks and produce output according to a strictly concave function of the labor input. When hiring new workers, employers incur fixed and variable costs of posting vacancies. Workers separate for exogenous reasons, and some employers layoff additional workers subject to fixed and variable firing costs. Employers choose layoffs and vacancies effectively, hires as well to maximize the present discounted value of profits. Workers do not search on the job. The following law of motion links employment changes and worker flows in CHW: e it = ( 1 q ) e 1 + η( U, V ) v it t t it l it, (2) 2 Other recent search-theoretic analyses with multi-worker firms include Elsby and Michaels (2008), Veracierto (2009), Fujita and Nakajima (2010), Schaal (2010) and Trapeznikova (2010). 6

9 where q is the quit rate, η U t, V ) is the job-filling rate, and v is the number of vacancies posted at the ( t it beginning of the period. 3 The job-filling rate derives from a standard matching function with constant returns to scale in the aggregate numbers of unemployed workers and vacant jobs. Period-t hires for the employer are h = η( U, V ) v. it t t it An employer in the CHW model operates in one of three regions each period, depending on the aggregate and idiosyncratic shock values: (i) positive vacancies and zero layoffs, (ii) zero vacancies and positive layoffs, or (iii) an inaction region with zero vacancies, no layoffs and no replacement hiring. Solving the model yields a stochastic equilibrium path for the cross-sectional distribution of employment changes, hires, quits and layoffs. The realized path depends in complex ways on the interaction of the aggregate and idiosyncratic driving forces and the key parameters of the revenue, cost and matching functions. Aggregate shocks shift the entire cross-sectional distribution of growth rates, while parameters governing adjustment costs and the variance of idiosyncratic shocks strongly influence its shape. Model-based outcomes exhibit an iron-link mapping of job flows to hiring, layoffs, and quits. Figure 3 depicts the iron link. It shows the relationship of the hiring, layoff, and quit rates to employer growth rates in the cross section. There is a mass point with no hiring or layoffs at growth rate q. To the right of q, employers post vacancies and hire new workers. Employers in this range have zero layoffs, and hiring rises one-for-one with increases in employment. To the left of q, employers engage in no hiring and the layoff rate rises one-for-one with the rate at which the employer contracts. Given the iron link relations illustrated in Figure 3, the cross-sectional distribution of employment growth rates fully determines aggregate hires and layoffs. This feature of the model has interesting implications. For example, modest recessions that shift the central tendency of the cross- 3 We use uppercase letters for aggregate quantities and lowercase letters for establishment outcomes. 7

10 sectional distribution from positive values to a value near zero produce large drops in aggregate hires with relatively little increase in aggregate layoffs. In contrast, a deep recession that shifts much of the mass in the cross-sectional distribution across the kink point in Figure 3 produces a large jump in layoffs. Thus, the highly nonlinear micro relations in Figure 3 imply differential responses of hires and layoffs to mild and deep recessions. Below, we treat the aggregate implications of cross-sectional worker flow relations in a more formal manner Relaxing the Iron Link There are several ways to relax the iron link feature of MP-style models. If we exogenously vary the quit rate in the CHW model, we obtain the implications shown in Figure 4. As the quit rate varies, the kink point moves and the layoff and hires relations shift accordingly. Specifically, a fall in the quit rate from q ( G0 ) to q G ) causes the cross-sectional hiring and layoff relations to shift rightward. That ( 1 is, the fall in the quit rate creates an environment where employers require fewer hires when expanding and more layoffs when contracting to achieve a given growth rate. More generally, if the quit rate varies systematically with labor market slack, then so do the cross-sectional relations for hires and layoffs. This insight applies in more complex models as well. Models with endogenous quit behavior deliver other interesting implications. As discussed in the introduction, the model of Faberman and Nagypál (2008) implies that quit rates decline with employer growth rates in the cross section. Workers are more likely to abandon struggling employers because they pay lower wages and cannot match outside offers from more profitable employers. Inspecting Figures 3 and 4, it is apparent that a nonlinear quit relation implies nonlinear relations for one or both of hires and layoffs as well. Somewhat paradoxically, weaker employers face greater needs for replacement hiring to offset, at least partly, a greater attrition rate. Thus, the model of Faberman and Nagypál can deliver a negative slope in the hires relation over some range. In the model of Barlevy 8

11 (2002), employers post fewer vacancies when aggregate conditions are weak. As a result, poorly matched employees encounter new job opportunities at a slower rate, and they quit less frequently. This implication of the Barlevy model provides an explanation for the downward shift in the quit rate in Figure 4 and the resulting shifts in the layoff and hiring relations. Models that feature learning about match quality as in Jovanovic (1979) yield additional insights. Under weak conditions, these models imply that employment relationships dissolve at a rate that declines with match duration. A newly hired worker is more likely to quit because he learns a job is not to his liking, or to be fired because his employer learns he cannot perform. Naturally, the proportion of new employees tends to rise with an employer s expansion rate. Thus, learning about match quality can cause quits and layoffs to rise with the growth rate of employment at expanding employers. Learning also leads to additional separations at shrinking employers when recent replacement hires do not work out. More generally, learning about match quality implies that worker flows beget further worker flows, as stressed by Hall (1995), Pries (2004), and Pries and Rogerson (2005). The lesson of this discussion is that the stark relations depicted in Figures 3 and 4 are too simple to fully capture the patterns in the data. Nevertheless, we think Figures 3 and 4 are useful starting points for two reasons. First, they provide a straightforward exposition of the links between worker flows and establishment-level growth in a prominent class of search models. We will investigate how closely the data conform to the relations exhibited in Figures 3 and 4. Second, our discussion provides guidance for formulating statistical models that relate worker flows to employer growth rates Implications for Aggregate Outcomes and the Relationship to Cross-Sectional Behavior To draw out the aggregate implications of the iron link relation and the other theoretical effects discussed above, express the aggregate hiring rate as 9

12 H f ( g) h ( g), (3) t = g t t where h t (g) is the mean hiring rate for establishments with growth rate g at time t, and f t (g) is the corresponding share of employment with growth rate g at t. This equation implies that movements in the aggregate hiring rate arise from changes over time in the cross-sectional relationship between hires and establishment growth rates, shifts in the growth rate distribution, and interactions between the two. Analogous remarks apply to separations, quits and layoffs. Of course, equation (3) is simply an accounting identity. Moving beyond the identity requires behavioral models of the micro relationships between worker flows and employer growth rates. Consider, for example, the CHW model with its time-invariant iron link between worker flows and establishment growth. Substituting the behavioral relations shown in Figure 3 into (3) yields the following expressions for aggregate rates of hires, layoffs and quits: H L t t = = g g f ( g) h( g), t f ( g) l ( g), and (4) t Q = f ( g) q = q t. t In the empirical work below, we fit, and relations using JOLTS micro data pooled over the 2001 to 2010 period. In doing so, we allow the quit rate to vary with employer growth, and we let the data freely determine the kink points, if any, in the cross-sectional relations. We also allow for hires in excess of job creation and separations in excess of job destruction. Even under these relaxed conditions, equations (4) preserve a key aggregate implication of the iron link: fluctuations in aggregate worker flows arise entirely from movements in the cross-sectional distribution of establishment growth rates. We will evaluate how well this implication describes the behavior of aggregate worker flows. g t 10

13 ~ More generally, the cross-sectional behavioral relations vary over time, and we write ht ( g ), ~ lt ( g ), and q ~ for the hiring, layoff and quit relations in period t. Proceeding as before, and replacing t ~ ~ the time-invariant relations in (4) with ht ( g ), lt ( g ), and q ~, yields the corresponding aggregate flow rates. When the cross-sectional behavioral relations vary over time, movements in the establishmentlevel growth rate distribution no longer suffice to fully account for fluctuations in aggregate worker flows. Nevertheless, we can still obtain informative characterizations of the aggregate flows in terms of ~ ~ statistical models for ht ( g ), lt ( g ), and q ~. Moreover, we can combine these models with data on f t (g) to construct synthetic measures of aggregate worker flows outside the period covered by JOLTS data. Figure 4 illustrates a case where the cross-sectional relations exhibit iron-link behavior in any given period, but they shift up and down over time with the aggregate quit rate. In this case, the behavioral relation for the hires rate takes the form t ~ ( g ) =, where X is a vector of ht indicators for aggregate labor market conditions and is a parameter vector that governs the size of the upward and downward shifts in the cross-sectional hires relations. Analogous remarks apply to the cross-sectional relations for separations, layoffs and quits. We investigate statistical models of this form and evaluate their ability to account for the cross-sectional and aggregate behavior of worker flows. We also investigate more flexible models that allow the shape of the cross-sectional relations to vary systematically with aggregate conditions. t 3. Data and Measurement We now turn to the description of the two data sources used and describe the key methodological concepts for the analysis. 11

14 3.1. Data Sources This study relies on two micro data sources, Business Employment Dynamics (BED) and the Job Openings and Labor Turnover Survey (JOLTS), both produced by the BLS. The BED contains longitudinally linked administrative records for all businesses covered by state unemployment insurance agencies virtually a census of nonfarm private business establishments. The data are quarterly and include employment and payroll for each establishment plus information on industry, location and whether the establishment belongs to a multi-unit firm. The BLS uses the BED to produce quarterly statistics on gross job creation and destruction from 1992, although micro data exist back to Our BED micro data run from 1990Q2 through 2010Q2. Data access restrictions preclude our use of BED data for Connecticut, Florida, Massachusetts, Michigan, Mississippi, New Hampshire, New York, Pennsylvania, and Wyoming. Time-series data on job creation and destruction rates generated from our version of the BED closely mimic the published series that cover all states. The JOLTS samples about 16,000 establishments each month and includes data on employment in the pay period covering the 12 th of the month. Establishments report hires, quits, layoffs, and other separations (deaths, retirements, and intra-firm transfers) over the course of the month. For quits, the establishments identify employees who left voluntarily (excluding retirements and intra-firm transfers). For layoffs, the establishments identify involuntary separations initiated by the employer. 5 The survey begins in December 2000 and covers the nonfarm economy. Our JOLTS micro data run from January 2001 through June For more details on the BED, see Spletzer et al. (2004). The BLS does not publish job flow statistics for because of issues related to administrative changes during that period. We follow Faberman (2008b) to address those changes. 5 The JOLTS survey instructions for layoffs and discharges include the following examples: layoffs with no intent to rehire, layoffs lasting more than 7 days, discharges resulting from mergers, downsizing, or closings, firings or other discharges for cause, terminations of permanent or short-term employees, terminations of seasonal employees (whether or not they are expected to return next season). For more details on the JOLTS, see Clark and Hyson (2001), Faberman (2008a) and Davis, Faberman, Haltiwanger, and Rucker (2010). Details on recent revisions to JOLTS methodology are available at 12

15 To construct quarterly worker flows at the establishment level, we require observations in all three months of the quarter. Dropping observations that violate this requirement reduces the number of establishments in our JOLTS sample by about 12 percent. This sample restriction produces slightly lower aggregate worker flow rates, but it does not alter their cyclical patterns. We also address other measurement issues related to the JOLTS data: timing differences in the measurement of worker flows and employment, the construction of sample weights at a quarterly frequency, and imputed worker flow rates for opening and closing establishments. The latter are covered by the BED but are not captured in the JOLTS sample frame. The appendix explains in detail how we address these matters. Our final sample contains over 277,000 establishment-quarter observations The Cross-Sectional Distribution of Establishment Growth Rates We compute employment growth as the difference between employment in the third month of the current quarter and the third month of the previous quarter, and we divide by the simple mean of current and previous employment to obtain a rate. (We use the same average of current and previous employment to compute worker flow rates.) This approach yields consistent aggregation and ensures that all growth rates are bounded, with entry and exit corresponding to values of +2 and -2. Given the comprehensive nature of the BED, we compute the cross-sectional distribution of employment growth rates directly from the micro data without need to adjust for sample weights. We focus on employmentweighted outcomes throughout the paper, unless noted otherwise. Figure 5 displays kernel density estimates of the establishment-level growth rate distribution in 2006Q1-2006Q4, a period of expanding aggregate employment, and 2008Q3-2009Q2, a period of sharp contraction. There is a clear leftward shift of the growth rate distribution from 2006 to the period. Table 1 reports summary statistics for the growth rate distribution in selected expansion and contraction periods. Establishments with steady employment levels over the quarter account for 13.9 to 13

16 16.1 percent of aggregate employment, depending on time period. Establishments that grow or shrink by more than 10 percent in the quarter account for 31 percent of employment in 1991, 29 percent in , and 26 percent in 2006 and 2008Q3-2009Q2. This evolution towards a more compressed growth rate distribution is also apparent in the behavior of aggregate worker flows Measuring Aggregate Worker Flows To measure aggregate worker flow rates, we combine JOLTS-based estimates of mean worker flow rates by growth rate bin with BED-based measures of the growth rate distribution. This method exploits identities of the sort described by equation (3). We use 195 bins that partition the full range of feasible growth rates, with narrower bins near the mode of zero growth. In defining bins, we allow for mass points at -200, 0 and 200 percent growth corresponding to exit, no change and entry. Letting w t (g) denote the mean worker flow rate for bin g in quarter t, we measure aggregate worker flow rates as W f ( g) w ( g). (5) t = g t t Because our JOLTS sample covers continuing establishments only, we do not have JOLTSbased worker flow rates for exits and entrants. Thus, we impute worker flow rates for entrants and exits when implementing (5). See the appendix for details. We prefer the aggregate worker flows computed according to (5) by combining BED and JOLTS data to the published JOLTS data, because our approach encompasses worker flows at exits and entrants, and because the BED provides a more accurate measure of the cross-sectional growth rate density. See Davis, Faberman, Haltiwanger and Rucker (2010) for additional discussion and analysis on this point. 4. Worker Flows in the Cross Section and Over Time Guided by the theoretical discussion in Section II, we specify statistical models of how worker flows vary with employer growth in the cross section and how the cross-sectional relations move over 14

17 time. We fit the models to establishment-level data and asses them in terms of cross-sectional fit, ability to replicate the time-series behavior of aggregate worker flows, and the marginal explanatory power of model-implied values for aggregate worker flows after conditioning on business cycle indicators. We also use the estimated statistical models to evaluate certain implications of the theoretical models Empirical Specifications Our first specification treats hires, separations, layoffs, and quits as time-invariant functions of establishment-level growth rates. Specifically, for each worker flow rate we regress the establishmentlevel observations on a vector of dummy variables for the 195 growth rate bins, w ( g) = α( g) + ε ( g) et et (6) where e indexes establishments, g indexes growth rate bins, and t indexes quarters. Fitted values in (6) given by wˆ D ( g) = ˆ α( g) describe the average cross-sectional worker flow relations in our JOLTS sample. This fixed cross section specification is consistent with the iron link feature of MP models, but it is flexible enough to accommodate learning about match quality as in Pries and Rogerson (2005) and the abandon-ship effect in Faberman and Nagypál (2008). Note that a strict MP-type iron link implies a perfect fit for (6). We examine R-squared values for fitted versions of (6) to quantify how closely the data conform to this implication. We also examine the shape of the fitted relationships to gauge whether and how closely they match the implications of various theoretical models. Recall that quit rates, and therefore hiring and separation rates, rise and fall in a pro-cyclical manner in the model of Barlevy (2002). To accommodate such behavior, our baseline specification relaxes (6) to let the cross-sectional relations shift up and down as functions of business cycle indicators: w et (g) = α (g) + β 1 G t + + β 2 G t + β 3 G t + β 4 JF t + ε et (g) (7) 15

18 where is the growth rate of aggregate employment, G = max{0, }, G = min{0, }, G t + t G t t G t G = G G t t t 1 is an accelerator term, and JF t is the job-finding rate calculated from Current Population Survey data on unemployment by duration. 6 To let the cross-sectional relations respond to business cycle conditions in more complicated ways, we extend the baseline specification by adding terms that involve interactions between the cycle indicators and the establishment-level growth rates. Specifically, we introduce a set of five indicator variables, I( g), for establishment-level growth rates less than or equal to -10 percent, greater than -10 percent but less than zero, exactly zero, positive but less than 10 percent, and greater than or equal to 10 percent. Thus, our flexible specification is given by w ( g) = α( g) + β G + + β G + β G + I ( g) δ G + I( g) δ JF + ε ( g). (8) et 1 t 2 t 3 t 1 t 2 t et Specification (8), unlike (7), allows the worker flow response to aggregate conditions to differ among establishments based on how rapidly they grow or shrink Worker Flows and Employer Growth in the Cross Section Figure 6 displays worker flow relations estimated from the fixed cross-section specification (6). The figure shows clear similarities to the theoretical relations in Figure 3. For example, the hiring relation exhibits a hockey-stick shape similar to the one implied by MP models. There are also clear departures from the implications of MP models. First, hires rise more than one-for-one with job creation to the right of zero. As discussed in Section II, this pattern suggests that a higher incidence of recently formed matches at more rapidly growing establishments leads to higher rates of learning about match quality, generating higher separation rates and more need for replacement hires. Second, hires occur at 6 We use standard methods to calculate the job-finding rate from data on unemployment by duration. For details, see Davis, Faberman, Haltiwanger, Jarmin and Miranda (2010). We use the published data from Current Employment Statistics to compute the growth rate of aggregate employment. 16

19 all growth rates, another piece of evidence that points to replacement hiring. Third, while there is a pronounced kink in the hiring relation, it occurs at zero growth rather than the mean quit rate. Finally, the hires rate declines with establishment growth rates over much of the range to the left of zero. This pattern is consistent with an abandon-ship effect in which more rapidly declining employers face a greater need for replacement hiring. Many of these same effects appear in the separations relation, which is nearly a mirror image of the hires relation. For example, separation rates rise with job creation rates to the right of zero, in line with the view that rapidly growing employers experience high turnover rates among recent hires. To the left of zero, separations rise more rapidly than job destruction, in line with the abandon-ship effect and a greater need for replacement hires at more rapidly shrinking employers. Turning to the breakdown of separations into quits and layoffs, several other patterns emerge. Layoffs rise sharply with job destruction, and they dominate the employment adjustment margin among rapidly shrinking employers. In contrast, quits account for a larger share of separations at establishments that shrink by less than 20 percent in the quarter and at growing establishments. In the appendix, we display a version of Figure 6 that covers a wider range of growth rates. The zoomed-out version of Figure 6 reveals even more clearly that quit rates top out at about 20 percent per quarter, and that layoffs are the primary margin of employment adjustment at rapidly contracting establishments. Quit and layoff rates, like the hiring rate, are smallest at employers with stable employment levels. 7 The patterns in Figure 6 differ from the cross-sectional relationships found by Abowd, Corbel, and Kramarz (1999) using establishment-level data for France. They find that hiring is the primary margin of employment adjustment at the establishment level, even for contracting establishments. This 7 For visual clarity, Figure 6 omits the cross-sectional relation for other separations (deaths, retirements, and intra-firm transfers). They are very small on average, amounting to about a half percent of employment, and somewhat greater at rapidly contracting establishments. 17

20 interesting point of contrast between their results and ours may reflect differences between France and the United States in the nature of labor adjustment. The topic warrants investigation in future work. Table 2 presents regression R-squared values for several specifications fit to the establishmentlevel JOLTS data. A potential concern about our specifications is that the 195-bin partition is too coarse to adequately approximate the underlying relationships of worker flows to employer growth rates. To investigate this concern, we augment each specification by adding 192 bin-specific slope terms. (There is no reason to introduce slope terms for the zero-width bins at -2, 0 and +2.) Comparing the first two columns in Table 2 reveals, however, that the bin-specific slope terms yield very modest gains in the regression fit of the fixed cross-section specification. For the sake of simplicity and parsimony, we drop the bin-specific slope terms in the rest of the paper. As discussed above, the strict iron-link feature of MP-type models implies a perfect fit for the fixed cross-section specification. We do not take this implication literally, because MP models deliberately abstract from many real-world features that play a role in worker flows. Still, it is interesting to ask how fully employer growth alone accounts for establishment-level variation in worker flows. According to Table 2, the fixed cross-section specification accounts for 54 percent of establishment-level variation in hiring rates and 51 percent for separation rates. It accounts for 47 percent of the variation in layoff rates and 16 percent of quit rate variation. These R-squared values fall well short of a perfect fit, but we are somewhat surprised by the extent to which the fixed cross-section model accounts for establishment-level variation in hires, separations and layoffs. In this regard, it is worth mentioning some of the many factors omitted from the model. In particular, it contains no controls for industry, employer size or age, wages, job tenure, education and other worker characteristics, the longer-term growth trajectory of the employer, local labor market conditions, and aggregate labor market conditions. Despite these omissions, this simple specification accounts for more 18

21 than half of the variation in hiring and separation rates across employers. Clearly, the employer growth rate is a major proximate determinant of worker flows. The remaining columns of Table 2 show the gains in fit from introducing cycle indicators and allowing for interactions between the cycle indicators and establishment-level growth rates. The baseline specification, which adds cycle indicators only, yields very modest gains in fit. The flexible specification, which also includes the interaction terms, improves the fit by several percentage points for each type of worker flow. Evidently, the relationship of worker flow rates to employer growth rates varies with aggregate labor market conditions. Table 2 also shows that even the flexible specification accounts for only about a quarter of the cross-sectional variation in establishment-level quit rates. The large role for establishment-level variation in quit rates conditional on own growth rate is an interesting finding, but it need not matter for movements in aggregate quit rates. To see this point, consider a multi-worker MP model with quit rates that vary exogenously among establishments. In particular, suppose the idiosyncratic component of the establishment-level quit rate is drawn from a common, time-invariant distribution. This assumption generates dispersion in worker flow rates among establishments with the same employment growth rate. Suppose, in addition, that the common component of quit rates varies with aggregate conditions as in our earlier discussion of Figure 4. Under these assumptions, there is dispersion in worker flow rates among establishments with the same employment growth rate, but common factors drive the average shape and location of the cross-sectional worker flow relations. For the analysis that follows, what matters is how well our statistical specifications capture variation in worker flow rates at the level of growth rates bins crossed with time periods. These binquarter outcomes provide building blocks for aggregation and the construction of our synthetic worker flow series. To obtain the bin-quarter outcomes, we compute the employment-weighted worker flow 19

22 rates for all 195 bins by quarter. To evaluate how well our statistical models capture the bin-quarter variation in worker flow rates, we estimate analogs of (6)-(8) at the bin-quarter level of aggregation. We summarize the main results of our regressions at the bin-quarter level here and report the full results in the appendix. The fixed cross-section specification accounts for 93 percent of the bin-quarter variation in the hiring rate and 92 percent for the separation rate. It accounts for 88 percent of the binquarter variation in layoff rates and 65 percent for the quit rate. The baseline and flexible specifications yield very modest improvements in fit for hires, separations and layoffs. The gains in fit are larger for the quit rate, with an R-squared value of 69 percent for the flexible specification. The flexible specification yields modest gains in fit relative to the baseline specification for quits and tiny gains for hires, separations and layoffs. That is, except perhaps for quit rates, the R-squared metric provides little evidence that the shapes of the cross-sectional relations vary systematically with the cycle indicators. The success of our statistical specifications in accounting for bin-quarter variation in worker flow rates bodes well for their ability to capture the behavior of aggregate worker flows Aggregate Behavior Implied by the Fixed Cross-Section Specification We now investigate how well our cross-sectional statistical models account for the behavior of aggregate worker flows. To start, we compare the actual worker flows to the aggregate flows implied by the fixed cross-section specification. To generate the implied flows, we replace the functions in (5) with wˆ D ( g) = ˆ α( g) obtained by fitting the fixed cross-section model to JOLTS data from 2001Q1 to 2010Q2. Plugging in the BED data for and the wˆ D ( g) functions yields the model-implied worker flow rates. 20

23 Figure 7 plots seasonally adjusted versions of the model-implied rates alongside the actual worker flow rates. 8 As seen in the figure, the layoff series implied by the fixed cross-section model captures much of the time-series variation in the actual layoff rate. The fixed cross-section model fares less well in replicating the actual hires and separations rate, and it completely fails to replicate the actual quit rates. In fact, the quit rate series implied by the fixed cross-section model is essentially a flat line for most of the period. It also predicts a rise in the quit rate during the downturn, which is nothing like the behavior of actual quit rates in this period. Table 3 quantifies the extent to which the fixed cross-section model replicates movements in the actual worker flow rates. (Like Figure 7, Table 3 includes results for other specifications that we discuss below). The first column of Table 3 reports the standard deviation of the actual worker flow rates. The remaining columns in the top panel report the root mean squared value of the difference between the model-implied and actual rates. For hires, the rate implied by the fixed cross-section model accounts for 36 percent of the time-series variation in the actual hires rate, computed as 100 times [1 (0.872/1.366)]. The correlation between the model-implied and actual hiring rate is For layoffs, the model-implied rate captures 41 percent of movements in the actual layoff rate, and the correlation between the two series is The model performs less well in replicating the behavior of the separations rate, and it captures very little of the time-series variation in the quit rate. We turn next to the cyclical behavior of the cross-sectional worker flow relations Shifts in the Cross-Sectional Relations over Time Before proceeding to the baseline and flexible specifications, we fit the fixed cross-section specification separately for three periods: 2001Q2 2003Q1 (a mild recession followed by a prolonged 8 We seasonally adjust the model-implied and actual worker flow rates using the Census X-11 procedure. We also drop the time series observations for 2001Q1 and Q2, because the JOLTS sample is much smaller in these early quarters about 2,700 observations per quarter as compared to more than 6,300 in other quarters. 21

24 jobless recovery ), 2006Q1 2006Q4 (an expansion period), and 2008Q3 2009Q2 (a deep recession). Each period covers four or eight consecutive quarters to ensure that seasonal effects do not drive the estimation results. For this exercise, we restrict attention to quarterly growth rates from -30 to 30 percent to focus on bins with ample observation counts. Establishments in this growth rate range account for about 90 percent of aggregate employment. Figure 8 displays the results. The layoff relation displays considerable stability over time. Rapidly shrinking establishments rely more heavily on layoffs (conditional on own growth) during the severe recession of , but the layoff relation is nearly time invariant throughout the rest of the growth rate range. In other words, there is a strong element of iron-link behavior in the cross-sectional layoff relation. In marked contrast, the quit relation shifts up and down as aggregate employment expands and contracts. Conditional on own-establishment growth, the quit rate is several percentage points lower in than in the other two periods. This pattern holds for the entire growth rate range and is most pronounced at rapidly shrinking establishments. Moreover, for shrinking establishments there is a clear ordering of the cross-sectional quit relations over the cycle: quit rates are highest in the 2006 expansion period, somewhat lower in the weak labor market early in the decade, and much lower in the severe recession of Extending our earlier metaphor, workers abandon ship at higher rates at more rapidly contracting employers, but they are much more likely to go down with the ship when stormy seas prevail. The separation relation largely inherits the cyclical behavior of the quit relation, although there are partly offsetting shifts in the underlying quit and layoff relations at rapidly shrinking establishments. Otherwise, the shifts in the separation relation look smaller only because the vertical scale covers a larger range. The cross-sectional hiring relation exhibits more stability over time than the separation 22